Skip to content

MPSBoost 1.0.0 Performance Report

This report records the final 1.0.0 performance and size evidence available on the release machine. It separates current-HEAD measurements from historical pre-v2/v3 benchmark baselines.

Release Machine

  • Date: 2026-07-23
  • Platform: Apple Silicon macOS
  • Python used for local audit: Python 3.13
  • Dataset cache policy: project-local ignored tests/real_world/data/ and tests/real_world/cache/
  • Wheel artifacts checked locally: dist/mpsboost-1.0.0-cp310-cp310-macosx_13_0_arm64.whl at 280K and dist/mpsboost-1.0.0-cp313-cp313-macosx_13_0_arm64.whl at 284K

Current-HEAD CPU Audit

These measurements use current source before the final 1.0.0 wheel rebuild. Peak RSS is converted from the process maximum resident set size and rounded to MiB.

Dataset Rows Features Estimator Train (s) Predict (s) Metric Score Model size
Iris 150 4 GradientBoostingClassifier 0.000965 0.000044 accuracy 0.973684 9,460 B
Breast Cancer 569 30 GradientBoostingClassifier 0.006503 0.000106 accuracy 0.930070 11,048 B
Diabetes 442 10 GradientBoostingRegressor 0.003520 0.000048 R2 0.424435 11,156 B
Digits 1,797 64 GradientBoostingClassifier 0.106488 0.001003 accuracy 0.740000 49,528 B
California Housing 20,640 8 GradientBoostingRegressor 0.079778 0.003057 R2 0.726251 33,000 B
Covertype subset 30,000 54 GradientBoostingClassifier 0.423550 0.006479 accuracy 0.473067 19,604 B

Peak RSS during this audit stayed under 495 MiB. The largest observed value was from the Covertype subset run after loading and splitting the cached dataset.

Historical MPS Baselines

The checked-in S4/S6 results remain useful for workload-shape guidance but are not final 1.0.0 speed claims. They were recorded before later v2/v3 implementation and cleanup work. Their role is to show the stable boundary: small or synchronization-heavy jobs can be slower on MPS, while wider and larger histogram workloads can benefit from Apple GPU execution.

The historical S6 large-wide regressor case reached 1.629x median speedup over the CPU oracle on the recorded M2 Ultra machine.

HIGGS Boundary

HIGGS is intentionally a local-file, opt-in performance-boundary dataset. The raw official dataset is multi-gigabyte scale and is not automatically downloaded into user environments or packaged in the wheel. The executable test remains in tests/real_world/test_external_opt_in_datasets.py and states the exact local file path required to run it.

1.0.0 Policy

  • Use device="auto" by default.
  • CPU is the correctness oracle and the preferred path for small, branch-heavy, and latency-sensitive workloads.
  • MPS is an acceleration backend for workloads large enough to amortize launch, transfer, and synchronization overhead.
  • Linux CPU/CUDA performance belongs to the selected external sklearn/XGBoost/CUDA stack and is reported as such.